Pre-sales

Does AI customer service increase sales? What research says

Field experiments and studies from Columbia, Harvard, Stanford and MIT on how AI customer service changes sales, and five rules for putting them to work.

· 8 min read · Editor: Max

A chatbot that answered shoppers' questions before they bought lifted an online retailer's sales more than five times as much as any other use of generative AI, the kind of software that writes its own replies, in randomized field experiments by a team that includes researchers from Columbia Business School. For a business weighing AI customer service, it is the clearest evidence yet that the payoff comes from answering buyers at the moment of choice, not from writing ads or product copy. Studies from Harvard, Stanford, MIT, Wharton and Temple University explain why it works, where it fails and where people still matter.

What happened when shoppers got an AI answer?

The main test ran in September and October 2023 on a leading cross-border online marketplace. More than 44,000 shoppers who asked about the platform's own products before buying were split at random. One group got what shoppers there had always got: an automatic note that service was unavailable, because the platform kept its agents for after-sale problems. The other got a chatbot that answered at any hour, in the shopper's language, about the exact product in front of them. It was one of seven experiments the team ran in 2023 and 2024, each placing the same technology somewhere else in the business.

Where the retailer used generative AIChange in salesChange in the share of shoppers who bought
A chatbot answering questions before purchase+16.3%+21.7%
Decoding what shoppers typed to find a product+2.9%+1.2%
Product descriptions+2.1%+1.3%
Marketing push messagesNo measurable change+3.0%
Advertising titlesNo measurable changeNo measurable change

The gain came from more people buying, not from bigger baskets: the chatbot's effect on order value was statistically insignificant. Return rates and ratings held steady, so the extra sales were not purchases buyers came to regret. Across the experiments, less experienced shoppers gained more than seasoned ones.

The team then set the chatbot against people. Shoppers it served spent 3.7% more than shoppers served by human agents, a gap within statistical noise. When the chatbot answered first and passed harder questions to a person, sales rose 25% over no pre-sale service, and shoppers spent 11.5% more than those served by human agents alone. The authors read that last comparison as a conservative lower bound on what generative AI adds to pre-sale support.

For a small business, the control group looks familiar: a contact form read on Monday, or a chat window that says the team is offline. Both leave a buyer where the marketplace's control group stood, holding a question at the moment of choice. That is where the leads you never see disappear: potential customers who were ready to leave a name and a need.

Will buyers trust a bot with a purchase?

Xueming Luo of Temple University's Fox School of Business and three co-authors tested that with more than 6,000 customers of a financial services company, in a study published in Marketing Science in 2019. When customers did not know they were talking to a bot, it sold as well as proficient human agents and four times as well as inexperienced ones. When the bot said what it was at the start of the conversation, purchase rates fell by 79.7%. Customers judged a disclosed bot less knowledgeable and less empathetic, although the bot itself was unchanged. Disclosing later in the conversation, and customers' earlier experience with AI, softened the penalty.

Hiding the bot is no answer. Since July 2019, California has made it unlawful to use a bot that misleads people about its artificial identity in order to encourage a sale. The authors trace the penalty to a perception of machines rather than to their competence, and a perception gives way to evidence. A first reply that names the right model, the price and the delivery date from the company's own pages is that evidence.

Most buyers want to help themselves first anyway. In a Harvard Business Review study of customer service from 2017, Matthew Dixon and three colleagues found that 81% of customers across industries try to take care of a matter themselves before they reach out to a live representative. A chat that answers from the company's own pages is self-service that talks back.

Why does a fast answer sell more?

Jan Carlzon, who turned Scandinavian Airlines around in the early 1980s, counted the moments that made or broke his company. In Moments of Truth, published in English in 1987, he wrote that each of the airline's 10 million customers dealt with some five employees a year, for an average of 15 seconds each time: 50 million moments a year that decided whether the airline succeeded or failed. A visitor's question in a website chat is one of those moments.

The cost of missing them was measured in a Harvard Business Review study from 2011. James Oldroyd, Kristina McElheran and David Elkington audited how 2,241 US companies answered a lead sent through their own websites. Among the companies that replied within a month, the first response took 42 hours on average.

In a companion analysis of 1.25 million leads, firms that answered within an hour had nearly seven times the chance of qualifying the lead, meaning a real conversation with a decision maker, as firms that answered even an hour later. Our review of speed-to-lead studies follows that line of work to the present.

Speed protects loyalty as well as sales. In a study of more than 75,000 customers, published in Harvard Business Review in 2010, Matthew Dixon, Karen Freeman and Nicholas Toman found that service interactions turn customers disloyal four times as often as they make them loyal. The damage comes from effort: contacting the company again, being transferred, repeating the same details.

What does AI change for the people on your team?

Economists Erik Brynjolfsson, Danielle Li and Lindsey Raymond, of Stanford and MIT, followed 5,172 customer support agents at a large software company after it gave them an AI assistant that suggested replies. In their 2025 paper in the Quarterly Journal of Economics, agents resolved 15% more issues per hour on average. Novices and lower-skilled agents improved in both speed and quality; the most experienced and highest-skilled saw small gains in speed and small declines in quality. Customers' messages turned more positive, and fewer of them asked for a manager.

A Harvard Business School working paper on 758 consultants at Boston Consulting Group drew the same line from another angle. On 18 tasks inside what the authors call the jagged frontier of AI's abilities, consultants with AI completed 12.2% more tasks, 25.1% faster, at more than 40% higher quality. On a task chosen to fall outside it, they reached the correct answer 19 percentage points less often than colleagues working without AI.

The practical line follows from both studies. Give AI what your pages already settle: prices, availability, delivery, how to order. Keep people on negotiation, exceptions and complaints, and make the switch with a clear human handover rule that passes the whole conversation along. Ethan Mollick of Wharton, a co-author of the consulting study, made "Always invite AI to the table" the first principle of his 2024 book Co-Intelligence, and set one limit beside it: legal and ethical barriers come first.

Why does a conversation beat a catalog?

In an upscale grocery store in California, Sheena Iyengar of Columbia and Mark Lepper of Stanford set up a tasting table with either 24 jams or six. Of the shoppers who stopped at the table, 30% bought jam when it offered six and 3% when it offered 24. A website with hundreds of products is the 24-jam table. A conversation that asks two questions and names two options turns it back into a table of six.

Theodore Levitt of Harvard Business School used to tell his students: "People don't want to buy a quarter-inch drill. They want a quarter-inch hole!" Clayton Christensen, also of Harvard Business School, built a theory on that idea: customers hire a product to make progress in their lives. A question typed into a website chat describes that progress in the buyer's own words.

For example, a tile shop gets a question at 11 p.m.: which tile works for a bathroom floor? The assistant asks for a name and a phone number so the team can confirm stock in the morning, then asks the size of the room and how much traffic the floor takes every day, then names two tiles from the shop's catalog pages with the price and the delivery time. The visitor goes to bed with an answer, and the shop opens with a lead.

What is a kept customer worth?

Sam Walton, who built Walmart, put it in one line: "There is only one boss. The customer." That boss, he added, can fire everybody in the company simply by spending money somewhere else.

Frederick Reichheld of Bain & Company and W. Earl Sasser of Harvard Business School showed in 1990 what that boss is worth. Cutting customer defections by 5% raised profits by 85% at a bank's branch system, by 50% at an insurance brokerage and by 30% at an auto-service chain. A buyer who gets a straight answer before the first purchase is the start of the relationship those numbers price.

Peter Drucker had opened with the same four words in 1954, in The Practice of Management: "There is only one valid definition of business purpose: to create a customer." A website that answers at midnight keeps creating customers while the team sleeps.

How do you put the research to work?

Five rules follow from these studies:

  1. Answer from your own pages. Prices, models and delivery dates come from what the company has published, never from guesswork.
  2. Answer at once, day and night, in the buyer's language. The marketplace's control group shows what silence costs.
  3. Narrow the choice. Ask one or two questions, then name two options, not twenty.
  4. Save every lead. A name and a way to reach the buyer let your team finish the sale the chat started.
  5. Hand over with the full thread. The person who takes over sees everything the buyer has said, so nobody repeats a word.

Then check the result with four numbers that matter, and compare tools with how to choose an AI chatbot for your website.

Axelion is built on these rules. It is an AI sales agent in your website chat: it answers in seconds, at any hour, in the visitor's language, from your own pages; it saves the visitor's name and contact as a lead; and your team continues the same conversation when a person is needed. Setup takes less than a day and needs no developer. The Free plan costs $0 and includes 3 conversations a month, and the plans page lists the rest.

Axelion answers in seconds, around the clock. It replies in your website chat, recommends the right product from your pages, saves every lead and hands complex cases to your team. WhatsApp and Telegram on Business. See it on your own website →

Related articles

AI chatbot pricing in 2026: what a small business pays

← All articles